Anthropic Unveils New AI Governance Architecture – Redefining Capital Rules for Global AI Companies

Dario Amodei, who once worked as Vice President of Research at OpenAI, co-founded Anthropic. With its Claude large language model, the company has become one of the world's leading AI companies.

When news spread that Anthropic might go public through an initial public offering (IPO), most people focused on its fast revenue growth, the abilities of its AI models, and its expected trillion-dollar valuation. Few paid attention to its unusual system of corporate control, which differs from the standard Silicon Valley model.

This is not surprising. For a long time, Silicon Valley technology companies preparing to go public have followed a familiar pattern. Most use a dual-class share structure, also called an A/B share structure, which gives founders more voting power than other shareholders.

Anthropic chose a different approach. It combined a Public Benefit Corporation (PBC) with a long-term independent trust.

Under this system, control of Anthropic belongs neither to its founders nor to its investors. Instead, some power is given to a group of outside trustees who own no company shares and receive no share of its profits. Former Federal Reserve Chairman Ben Bernanke is among these trustees.

This approach challenges a long-standing belief in financial markets: that public companies should put shareholder returns first.

But if AI safety and the public good come before financial returns, how should investors in the stock market value the company?

Does this mean investors' shares are worth less because they have fewer rights? This is not just a theoretical question about the law. It is a real challenge that Wall Street will face when Anthropic goes public.

What happens next could also offer a useful example for other large AI model companies that plan to go public in the future.

How the two-part structure works: The trust holds real power

In the US stock market, most technology companies use a traditional business structure known as a C corporation, or C Corp.

For these traditional companies, the law generally requires managers and boards of directors to focus on increasing shareholder value when making major decisions. If managers give up large short-term profits to protect safety or reduce social harm, shareholders may be able to sue the board and seek compensation, depending on the circumstances.

A Public Benefit Corporation (PBC) is also a for-profit business. It can raise money in the usual ways, issue shares, go public, and pay dividends in the future. It competes in the market much like an ordinary company.

The main difference is that it changes the board of directors' legal responsibilities.

Anthropic's governance system combines two parts. The first is the Public Benefit Corporation, which serves as the company's legal structure. The second is a long-term independent trust that holds actual control. Together, they create a special system of checks and limits on power.

Under PBC rules, managers do not have to put shareholder returns above everything else. They can also legally consider public safety and the effects of their decisions on society.

AI companies choose very different governance systems: Anthropic takes a middle path

Anthropic was not the first company to use this type of business structure. Outdoor clothing brand Patagonia, crowdfunding platform Kickstarter, and ice cream brand Ben & Jerry's all adopted the PBC structure early on.

Among the world's leading AI companies, three have taken very different approaches to corporate structure: OpenAI, Anthropic, and xAI.

OpenAI's main business entity, which handles its commercial activities, is also registered as a PBC. But OpenAI uses a special structure: a nonprofit foundation sits at the top and controls the organization, while a for-profit company below it runs the business. This arrangement is intended to limit the influence of investors seeking profits.

Anthropic has chosen more of a middle path than OpenAI. Both companies have systems designed to limit the power of capital, but their basic structures are quite different.

OpenAI has a nonprofit foundation at the top. Anthropic, by contrast, remains a for-profit Public Benefit Corporation (PBC).

Anthropic separates control of the company from its normal business operations and gives that control to outside trustees through an independent trust. OpenAI's foundation directly owns shares, while Anthropic's trust owns no shares and instead holds voting control.

Elon Musk founded xAI in 2023, originally aiming to compete with OpenAI. The company briefly registered as a PBC but later abandoned that structure and returned to the ordinary for-profit company model.

xAI removed this extra legal commitment to a public mission, so its decisions focus on business growth. It has more freedom in how it handles investment and expansion, without limits imposed by this type of governance system. Its safety standards depend on the company policing itself, rather than on firm legal rules or oversight by a trust.

Anthropic sits between these two approaches. It can still raise money through normal market channels and is not subject to the same nonprofit structure as OpenAI. At the same time, its independent trust turns its safety mission into a firm limit on corporate decisions, making it less likely that short-term pressure from investors will override long-term risk management.

Anthropic is trying to balance two goals that can be difficult to combine: freedom to grow its business and responsibility to the public because of the power of AI technology.

These different approaches reflect a deeper disagreement across the industry. Are very large AI models simply ordinary commercial products, or are they more like essential infrastructure that can create serious public risks? Different answers to this question have led to very different corporate structures.

Why do large AI model companies care so much about corporate responsibility?

Large AI models can create risks that spread beyond the companies that build them. If companies rush to sell and expand their products without enough care, they may create a range of social risks.

In an ordinary business, investors could pressure managers to release more model features or launch risky products to improve short-term financial results. A PBC gives the board legal room to put AI safety ahead of short-term profits.

But Anthropic believes a PBC alone is not enough. It can give the board legal protection, but it may not stop shareholders or founders from changing the company's direction when the chance to make huge profits becomes too tempting. To address this problem, Anthropic also created the LTBT, a long-term independent trust, to add a second layer of control.

The PBC is only the company's legal structure. By itself, it cannot fully limit the power of investors. The LTBT, the long-term independent trust, holds the real power. This is also the most controversial part of the system.

The trust's members are experts chosen from outside the company. They do not own Anthropic shares, cannot profit from an increase in the company's share price after it goes public, and do not receive company dividends. Their pay is only meant to cover the time they spend on their duties.

The trustees have two main powers. First, they can replace most members of the board of directors, giving them major influence over who leads the company. Second, the company must notify the trust five days in advance of major transactions, such as taking on unusually large amounts of debt, selling more than half of the company's shares, or transferring core AI models when the transfer meets certain thresholds.

After reviewing the relevant rules, many scholars of business law have pointed out that this five-day notice is more than a simple update. In practice, it can give the trust an informal power to block a deal.

If the trust believes a transaction could create safety risks, it can try to stop the deal by appointing or removing key people and applying continued pressure.

This is very different from the A/B share model. In that model, founders keep control by holding shares with extra voting power, so authority stays within the company. Anthropic instead gives this key power to a third party that is independent of both its shareholders and its founding team.

Long-term advantages and an unavoidable discount in value

Anthropic's system effectively means that shareholders give up some of their usual rights from the start.

This structure has led to two opposing views among investment firms. Investors who buy shares before a company goes public are also much more willing to accept this arrangement than investors who trade its shares on the public market.

Those who support the structure have a clear argument. The trust acts as an independent watchdog, which can prevent the power struggles and internal conflicts between founders and investors that are common in Silicon Valley companies. This could help the company stick to its long-term plans.

As governments around the world tighten AI rules, a governance system with built-in safety limits can also give Anthropic an advantage in meeting regulations. Investors would find it harder to pressure the team into taking technical risks just to improve quarterly results. Long-term work, such as research and development and safety testing, could continue to receive funding.

Investors in private markets are more willing to pay for this long-term promise. But concerns among public-market investors are also realistic, and they are a major source of uncertainty about the stock's future price.

The main problem is that shareholders have fewer rights by design. Investors buy shares, but they may have little influence over deals that could change the company's future.

A group of trustees who do not answer to shareholders controls decisions on major matters. If the trust blocks a merger or licensing deal that could bring in huge profits because of safety concerns, shareholders may have no choice but to accept the decision, with no effective way to challenge it.

Traditional investment research can estimate revenue, profit margins, and the strength of competitors. But it is difficult for the market to build a valuation model around the personal judgments of the trustees.

This creates a problem that may be hard to avoid. Anthropic has leading technology, a fast-growing business, and many features of a high-growth company. Yet the shares do not provide investors with all the usual rights. The market may eventually value the company less because those rights are limited.

Private-market investors may accept this discount because they are betting on long-term growth. But many public-market investors work within clear time frames for judging their investments, so they are naturally more cautious about uncertain risks linked to corporate control.

This difference in how investors view the company could continue to shape its stock price for a long time after Anthropic goes public.

A governance experiment for the whole industry, with more tests ahead

Looking further ahead, the importance of this issue goes beyond whether Anthropic's own IPO succeeds.

In the short term, interest in AI is strong, and the limited supply of leading AI companies available to investors may overshadow arguments about corporate control. As a result, the IPO is likely to go ahead smoothly, with plenty of demand for shares.

In the medium term, stock price changes after the IPO and the completion of major deals are likely to bring repeated debate about the trust's broad powers and the lack of clear rules. If the trustees make decisions that directly conflict with investors' interests, shareholder lawsuits, changes in institutional investors' holdings, and new assessments of the company's value are all possible outcomes.

In the long term, the example Anthropic sets could have a major impact. Governments around the world are likely to keep tightening AI safety rules. A model that relies only on companies policing themselves and founders voluntarily following safety principles may gradually become less effective.

If Anthropic's system works well and maintains a balance between business growth and risk management, it could become a new model for leading AI companies that build large models.

For a hundred years, capital markets have largely operated on the idea that shareholders should come first. Wave after wave of new technology has changed products and business models, but this basic approach to corporate control has rarely faced a serious challenge.

AI is different. The risks it creates can spread widely, and its effects on society go beyond those of the internet and semiconductor industries. The old rules of business and investment may not fit this new technology.

We are now seeing a slow process of adapting institutions to this new reality. For now, the industry has no standard answer to whether investors should give up some of their power to protect public safety, or how companies should balance business profits against long-term risks.

Through its IPO, Anthropic is testing a new approach for the entire industry. The final outcome will not only shape the story investors tell about Anthropic's stock price. It could also influence how future AI giants organize themselves when they enter public markets.

This article comes from the WeChat Official Account "Wall Street Tech Eye". The author is Du Jie, and the editor is Sun Congcong. 36Kr published the article with permission.